If you have a tiny model (robotics, RL) cpu-overhead bound, avoid frequent calls to eval() or train() in eager mode, or model.parameters() or anything that goes through your model. Prefer cached versions of these calls.

Comments